Developing Bayesian network inference algorithms to predict causal functional pathways in biological systems

Network inference algorithms are powerful computational tools for identifying potentially causal interactions among variables from observational data. Bayesian network (BN) inference algorithms hold particular promise in that they can capture linear, nonlinear, combinatorial, stochastic, and other types of relationships among variables across multiple levels of biological organization. However, challenges remain when applying these algorithms to limited amounts of experimental data collected from biological systems, and when dealing with the computational complexity. In addition, verifying the results from the algorithms is problematic, since biological knowledge of the causal interactions is limited and difficult to obtain.; In this dissertation, I use a simulation approach to advance BN inference algorithms, especially in the context of limited amounts of biological data. I develop an influence score and a data interpolation strategy to increase the information recovered from BN inference algorithms, and to reduce false predictions when faced with limited data. I prove the approximate lower bounds on the amounts of data required for accurate network recovery. I make improvements to the BN model in order to achieve computational efficiency and accuracy, including methods for finding an optimal directed acyclic graph, adding informative priors, and employing higher-order Markov frameworks for dynamic Bayesian networks (DBN).; After developing and testing the BN inference algorithms with simulated data, I apply them to two different types of experimentally measured biological data to solve questions in different biological systems. First, I apply my DBN inference algorithms on time-series gene expression data from Novartis Pharmaceuticals, to study compound-related effects at the molecular level. Second, I apply the DBN algorithms on electrophysiological data collected from songbird brain when birds are hearing or singing, to study neural information flow in the songbird auditory and vocal systems. Biological verifications for the inferred networks, where possible, demonstrate that the algorithms recover highly accurate networks. The recovered networks further make predictions about the interactions in those biological systems. The set of BN inference algorithms developed in this dissertation allow investigators to take microarray gene expression and microelectrode array data, and successfully generate genetic regulatory and neural information flow networks. This Bayesian approach is novel to the field of neuroscience. By applying to both electrophysiology and gene expression data from the brain, the recovery results from our algorithms can provide insights into how the brain perceives the world and produces behavior. The biologically plausible results from our applications suggest that the BN inference algorithms are potentially powerful tools to decipher casual pathways involved in complex biological functions.